Olson tackles watch

[gumby013]

I wouldn't put a ton of confidence in the notion that what I'm doing is correct (it's not back tested, assumptions are borrowed from other places that may not apply, and no one's paycheck is contingent on me being even reasonably correct) but these guys help me get it done:

average()
stdev.s()
norm.inv()
rand()

I googled the cone of uncertainty graph.

Would stdev.p be valid in calculating the cone? I use that one a lot more in my work.

[bisonaudit]

I hate to point this out but you need to put these numbers to at least one decimal point, because right now you are saying Grant is guaranteed to get to #5. He could have some kind of season ending injury in practice today that would keep him at 310. Considering he has never missed a game, it's a very small percentage, but it is still there.

The model knows nothing about injury risk other than what shows up indirectly in his historical tackles by game. All models are an approximation of reality and you are absolutely correct, injury risk is a feature of reality that may not be approximated very well in this model.

That said there are two games in the data set where Grant had 0 tackles and w/ a weighted average around 8 and a standard deviation around 6 the model does incorporate zero tackle games as a possibility (6 in the first 7 projections of the rest of the season that I'm looking at now). I run just 1,000 sims at a time. If I pound the F9 key it's coming up either 999 or 1,000 out of a 1,000 to get to 331 so depending on which model run I screen cap for you all it may still have kicked out a flat 100%.

Even if he had been hurt and missed say 3 games in a row; the model wouldn't know that. It would just see 3 independent zeros and incorporate them into the average and standard deviation and do it's stupid normal distribution calculation. It's a limitation.

It's inexpertly borrowed from some of some stuff Hollinger (basketball) and before him James (baseball) did to project career statistics (Kobe was about 50/50 to catch Kareem before his Achilles exploded). Those incorporated personal injury history just by looking at the stats the way this thing does (indirectly) which makes some sense over a number of very long seasons. They also contain assumptions on aging and productivity curves that aren't applicable to a 4 year collegiate career. Those curves indirectly incorporate the overall risk of injury in the various sports they're tuned for but I striped them out.

[bisonaudit]

Would stdev.p be valid in calculating the cone? I use that one a lot more in my work.

My theory for using stdev.s is that the population is all career games (including those that haven't been played yet) and the data I've got so far constitutes the sample. If you're just trying to describe what's happened so far (which I think is the goal of your rankings) then stdev.p seems right. That's just my theory though. I took one semester of stats at NDSU and the rest of this I've just cobbled together after discovering sabermetrics.

[bisonaudit]

"cone of uncertainty" love it

Stole that from the National Hurricane Center.

[CAS4127]

The model knows nothing about injury risk other than what shows up indirectly in his historical tackles by game. All models are an approximation of reality and you are absolutely correct, injury risk is a feature of reality that may not be approximated very well in this model.

That said there are two games in the data set where Grant had 0 tackles and w/ a weighted average around 8 and a standard deviation around 6 the model does incorporate zero tackle games as a possibility (6 in the first 7 projections of the rest of the season that I'm looking at now). I run just 1,000 sims at a time. If I pound the F9 key it's coming up either 999 or 1,000 out of a 1,000 to get to 331 so depending on which model run I screen cap for you all it may still have kicked out a flat 100%.

Even if he had been hurt and missed say 3 games in a row; the model wouldn't know that. It would just see 3 independent zeros and incorporate them into the average and standard deviation and do it's stupid normal distribution calculation. It's a limitation.

It's inexpertly borrowed from some of some stuff Hollinger (basketball) and before him James (baseball) did to project career statistics (Kobe was about 50/50 to catch Kareem before his Achilles exploded). Those incorporated personal injury history just by looking at the stats the way this thing does (indirectly) which makes some sense over a number of very long seasons. They also contain assumptions on aging and productivity curves that aren't applicable to a 4 year collegiate career. Those curves indirectly incorporate the overall risk of injury in the various sports they're tuned for but I striped them out.

My theory for using stdev.s is that the population is all career games (including those that haven't been played yet) and the data I've got so far constitutes the sample. If you're just trying to describe what's happened so far (which I think is the goal of your rankings) then stdev.p seems right. That's just my theory though. I took one semester of stats at NDSU and the rest of this I've just cobbled together after discovering sabermetrics.

Now you are just showing off!!

[gumby013]

My theory for using stdev.s is that the population is all career games (including those that haven't been played yet) and the data I've got so far constitutes the sample. If you're just trying to describe what's happened so far (which I think is the goal of your rankings) then stdev.p seems right. That's just my theory though. I took one semester of stats at NDSU and the rest of this I've just cobbled together after discovering sabermetrics.

Interesting. I was wondering because in my actual job, I utilize cones of uncertainty frequently. They all use stdev.p in their generation though. I'll have to explore stdev.s a bit more. Thanks!

Cone of Uncertainty.jpg

[bisonaudit]

Interesting. I was wondering because in my actual job, I utilize cones of uncertainty frequently. They all use stdev.p in their generation though. I'll have to explore stdev.s a bit more. Thanks!

Cone of Uncertainty.jpg

Interesting. Stdev.s should result in a marginally wider cone. Someone on here must have access to a stats professor who could perhaps further enlighten us both.

[KSBisonFan]

I have no idea what the gibberish and graphs in this thread say but let me be the first to congratulate Grant Olson on breaking the tackles record Saturday against ISU-B!

[reformedUNDfan]

are you people doing stats stuff with excel?

[bisonaudit]

are you people doing stats stuff with excel?

It would be a stretch to call what I do with excel a formal statistical analysis.

While you wouldn't want the results of a drug trial or the existence of the Higgs Boson riding on a statistical analysis performed in excel. I think it can handle an entertaining projection of a 1-AA middle linebacker's career tackles or even assist in identifying quality plays against the spread (for entertainment purposes only).

Besides, user error (of statistics not excel) is a much bigger problem at the level of sophistication we're talking about here than any deficiencies excel may have at the margin. http://thebiglead.com/2013/10/24/esp...led-miserably/